Fix issues from bad rebase

alpaka-group · Aug 12, 2024 · 4ad8bae · 4ad8bae
1 parent c6c12fd
commit 4ad8bae
Show file tree

Hide file tree

Showing 2 changed files with 30 additions and 2 deletions.
diff --git a/include/alpaka/kernel/TaskKernelGpuCudaRt.hpp b/include/alpaka/kernel/TaskKernelGpuCudaRt.hpp
@@ -11,9 +11,9 @@
 
 namespace alpaka
 {
-    template<typename TAcc, typename TDev, typename TDim, typename TIdx, typename TKernelFnObj, bool TCooperative, typename... TArgs>
+    template<typename TAcc, typename TDim, typename TIdx, typename TKernelFnObj, bool TCooperative, typename... TArgs>
     using TaskKernelGpuCudaRt
-        = TaskKernelGpuUniformCudaHipRt<ApiCudaRt, TAcc, TDev, TDim, TIdx, TKernelFnObj, TCooperative, TArgs...>;
+        = TaskKernelGpuUniformCudaHipRt<ApiCudaRt, TAcc, TDim, TIdx, TKernelFnObj, TCooperative, TArgs...>;
 } // namespace alpaka
 
 #endif // ALPAKA_ACC_GPU_CUDA_ENABLED
diff --git a/include/alpaka/kernel/TaskKernelGpuUniformCudaHipRt.hpp b/include/alpaka/kernel/TaskKernelGpuUniformCudaHipRt.hpp
@@ -469,6 +469,34 @@ namespace alpaka
                 return kernelFunctionAttributes;
             }
         };
+
+        //! The CUDA/HIP get max active blocks for cooperative kernel specialization.
+        template<typename TAcc, typename TKernelFnObj, typename TApi, typename TDim, typename TIdx, typename... TArgs>
+        struct MaxActiveBlocks<TAcc, DevUniformCudaHipRt<TApi>, TKernelFnObj, TDim, TIdx, TArgs...>
+        {
+            ALPAKA_FN_HOST static auto getMaxActiveBlocks(
+                TKernelFnObj const& kernelFnObj,
+                DevUniformCudaHipRt<TApi> const& device,
+                alpaka::Vec<TDim, TIdx> const& blockThreadExtent,
+                alpaka::Vec<TDim, TIdx> const& threadElemExtent,
+                TArgs const&... args) -> int
+            {
+                auto const blockSharedMemDynSizeBytes
+                    = getBlockSharedMemDynSizeBytes<TAcc>(kernelFnObj, blockThreadExtent, threadElemExtent, args...);
+
+                int numBlocksPerSm = 0;
+                TApi::occupancyMaxActiveBlocksPerMultiprocessor(
+                    &numBlocksPerSm,
+                    alpaka::detail::
+                        gpuKernel<TKernelFnObj, TApi, TAcc, TDim, TIdx, remove_restrict_t<std::decay_t<TArgs>>...>,
+                    blockThreadExtent.prod(),
+                    static_cast<std::size_t>(blockSharedMemDynSizeBytes));
+
+                auto multiProcessorCount = trait::GetAccDevProps<TAcc>::getAccDevProps(device).m_multiProcessorCount;
+
+                return numBlocksPerSm * multiProcessorCount;
+            }
+        };
     } // namespace trait
 } // namespace alpaka